A lightweight method to customize a frozen language model for specific tasks without retraining the entire model.